GenericAgent (GA)

A Token-Efficient Self-Evolving LLM Agent via Contextual Information Density Maximization

Advantage AI Agent Lab (A³ LAB) · Shenzhen Aquaintelling Technology × Fudan University

---

GA is a self-evolving LLM agent that matches or beats Claude Code / OpenClaw on hard tasks while spending ~1/6 the tokens. It uses just 9 atomic tools, a 92-line core loop, and gets cheaper the more you use it (−89.6% tokens over 9 rounds).

This repo hosts the Technical Report (PDF · arXiv) and all evaluation datasets. For the runnable agent, see the source-code repo: GenericAgent.

Contents

• Contents
• News
• Highlights
• How It Works
• Evaluation Results
  • 1. Task Completion & Token Efficiency
  • 2. Tool-Use Efficiency
  • 3. Memory System Effectiveness
  • 4. Self-Evolution Capability
  • 5. Web Browsing Capability
• Repository Layout
• Citation

News

• 2026.04 — Featured on 机器之心 (jiqizhixin).
• 2026.04 — Technical Report v1.0 released on arXiv.

Highlights

• 🪶 9 atomic tools, not 50+ — capability through composition (vs. 53 in Claude Code).
• 📉 ~1/6 the token cost at matched-or-better success (188.8k vs 537.4k on the 5-task long-horizon benchmark).
• 🔁 Self-evolving — −89.6% tokens, −78.2% runtime, −84.4% LLM calls over 9 rounds, no prompt tuning.
• 📚 No vector DB — beats embedding retrievers (Mem0, A-MEM) on LoCoMo with pure hierarchical memory.
• 🛡️ No context explosion — 2,298-token prompt after 20 skills (~10–20× smaller than baselines).
• 🌐 Web-ready — 3× the BrowseComp-ZH score at ~1/6 the tokens vs. OpenClaw.
• ⚙️ Minimal — 92-line core loop, ~3.3k LoC total (vs. ~530k for OpenClaw).

How It Works

GA runs a unified agent loop and maximizes context information density across four mechanisms:

Mechanism	Idea
Minimal Atomic Toolset	9 primitives across 5 capability classes; broad capability via composition, not enumeration.
Hierarchical Memory	4 layers (L1 index → L2 facts → L3 SOPs → L4 archive); only the compact L1 stays in-prompt, the rest is retrieved on demand.
Self-Evolution	Reflection compresses verified trajectories into SOPs → executable code; triggered by the memory system, not the user.
Context Compression	Truncation, periodic tag-level compression, FIFO eviction past 60% budget — targeting a <30k token budget.

Emergent capabilities built on top: subagent dispatch (GA CLI as a sub-process), Reflect Mode (event-driven watchdog + cron-style tasks), and autonomous exploration over a persistent skill tree.

Evaluation Results

GA is evaluated along five dimensions against Claude Code, OpenAI Codex, and OpenClaw under Claude Sonnet 4.6 / Opus 4.6, GPT-5.4, and MiniMax M2.7 backbones.^† Full details in the Technical Report.

^† Model version names follow our internal evaluation pipeline; see §4.1.1 of the report for exact backbone configs.

1. Task Completion & Token Efficiency

Efficiency = Accuracy / Total Tokens (M).

Benchmark	Agent	Backbone	Accuracy	Total Tokens	Efficiency
SOP-Bench	GA	Claude Sonnet 4.6	100%	2.08M	0.48
	OpenClaw	Claude Sonnet 4.6	100%	2.64M	0.38
	Claude Code	Claude Sonnet 4.6	85%	1.25M	0.68
	GA	MiniMax M2.7	90%	0.92M	0.97
	OpenClaw	MiniMax M2.7	95%	2.96M	0.32
Lifelong AgentBench	GA	Claude Sonnet 4.6	100%	0.24M	4.15
	OpenClaw	Claude Sonnet 4.6	70%	1.45M	0.48
	Claude Code	Claude Sonnet 4.6	75%	0.81M	0.92
	GA	MiniMax M2.7	90%	0.42M	2.12
	OpenClaw	MiniMax M2.7	70%	1.22M	0.57
RealFin-Benchmark	GA	Claude Sonnet 4.6	65%	0.11M	5.70
	Claude Code	Claude Opus 4.6	60%	0.31M	1.95
	Claude Code	Claude Sonnet 4.6	55%	0.24M	2.31
	OpenClaw	Claude Sonnet 4.6	35%	0.25M	1.39
	Codex	GPT-5.4	60%	0.89M	0.67

📌 On Lifelong AgentBench, GA hits 100% accuracy with 27.7% of Claude Code's input tokens and 15.5% of OpenClaw's.

2. Tool-Use Efficiency

GA exposes 9 atomic tools, vs. 53 in Claude Code (20 base + 33 conditional) and 18 factories in OpenClaw.

Long-horizon complex tasks (5 tasks, Claude Sonnet 4.6):

Agent	Success	Total Tokens	Time (s)	Requests	Tool Calls
GA	100.0%	188,829	220.8	11.0	12.8
Claude Code	100.0%	537,413	320.8	32.6	22.6
OpenClaw	80.0%	633,101	183.1	15.0	16.6

📌 GA matches Claude Code's success while using 35.1% of its tokens, 33.7% of its requests, 56.6% of its tool calls.

_{Tool-use efficiency radar — GA leads on token, request, and tool-call axes while preserving quality.}

3. Memory System Effectiveness

(a) Condensed memory ablation — SOP-Bench (dangerous goods):

Configuration	Memory Size (tokens)	Task Success Rate
No-Memory	0	13.87%
Full-Memory	575	52.44%
Redundant-Memory	288	66.48%
Condensed Memory	165	66.48%

(b) Long-term factual memory — LoCoMo (no embedding model, no vector DB):

System	Multi-Hop F1	Temporal F1	Open-Domain F1	Single-Hop F1
Mem0	39.32	50.03	18.32	40.32
A-MEM	29.03	46.83	13.11	44.68
OpenClaw	21.43	22.56	9.56	23.44
GA	43.33	52.23	20.41	45.69

(c) Context-explosion stress test — full prompt length after installing 20 skills:

System	Full Prompt Length
Claude Code	22,821 tokens
Codex	23,932 tokens
OpenClaw	43,321 tokens
GA	2,298 tokens

📌 GA's prompt stays ~10–20× smaller after extensive skill expansion — idle memory is kept strictly off-prompt.

4. Self-Evolution Capability

Nine-round longitudinal study (LangChain GitHub research task). Stage transitions are triggered autonomously by the memory system.

Round	Stage	Time	LLM Calls	Total Tokens
#1	Initial run	7m30s	32	222,203
#2	SOP optimization	4m19s	12	66,341
#3	SOP optimization	2m53s	8	49,825
#4	SOP optimization	2m29s	9	51,758
#5	SOP optimization	2m50s	7	35,536
#6	Codified SOP	2m24s	6	25,762
#7	Codified SOP	1m41s	5	23,014
#8	Codified SOP	1m35s	5	22,689
#9	Codified SOP	1m38s	5	23,010

📌 Round #1 → #9: −89.6% tokens, −78.2% runtime, −84.4% LLM calls — savings come from removing whole reasoning loops, not just shorter responses.

Across 8 web tasks (3 runs each), GA cuts tokens by 61.0%–92.4% (overall −79.3%); OpenClaw shows no comparable convergence.

_{GA converges to a stable low-cost regime across eight web tasks; OpenClaw does not.}

5. Web Browsing Capability

Both GA and OpenClaw use Claude Opus 4.6.

Benchmark	# Tasks	Eval Protocol	GA Score	OpenClaw Score	GA Tokens	OpenClaw Tokens
WebCanvas	12	Automatic	0.834	0.722	0.18M	0.71M
BrowseComp-ZH	10	LLM-as-Judge	0.600	0.200	0.47M	1.31M
Custom Tasks	22	Human + LLM	0.577	0.500	0.26M	0.76M

📌 On multi-hop BrowseComp-ZH, GA triples OpenClaw's score (0.60 vs. 0.20) at ~1/6 the tokens; 2.9×–3.9× token reduction across all three benchmarks.

Repository Layout

GA-Technical-Report/
├── GA_Technical_Report.pdf       # Full technical report
├── assets/                       # Figures
└── datasets/                     # Evaluation datasets (each with its own README)
    ├── lifelong_agentbench/
    ├── locomo/
    ├── realfin_benchmark/
    ├── sop_bench/
    ├── tool_efficiency_benchmark/
    └── web_browsing/

Each benchmark folder ships its own README with task setup, evaluation protocol, and reproduction notes.

Citation

@misc{liang2026genericagenttokenefficientselfevolvingllm,
      title={GenericAgent: A Token-Efficient Self-Evolving LLM Agent via Contextual Information Density Maximization (V1.0)}, 
      author={Jiaqing Liang and Jinyi Han and Weijia Li and Xinyi Wang and Zhoujia Zhang and Zishang Jiang and Ying Liao and Tingyun Li and Ying Huang and Hao Shen and Hanyu Wu and Fang Guo and Keyi Wang and Zhonghua Hong and Zhiyu Lu and Lipeng Ma and Sihang Jiang and Yanghua Xiao},
      year={2026},
      eprint={2604.17091},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.17091}, 
}

_{© 2026 Advantage AI Agent Lab (A³ LAB). Released alongside the GenericAgent system at github.com/lsdefine/GenericAgent.}